Reproduce “Diffusion-Scheduled Denoising Autoencoders for Anomaly Detection in Tabular Data” on OpenResearch

About this paper

Timur Sattarov, Marco Schreyer, Damian Borth

Researchers from Deutsche Bundesbank, Swiss Federal Audit Office, and University of St.Gallen developed Diffusion-Scheduled Denoising Autoencoders (DDAE) and a contrastive variant (DDAE-C) for anomaly detection in tabular data. This framework leverages diffusion model noise scheduling and achieves state-of-the-art results in semi-supervised settings, showing a 65.1% PR-AUC improvement over traditional DAEs in unsupervised anomaly detection on the ADBench benchmark.

Key insights

• Diffusion-based noise scheduling provides a substantial enhancement over fixed-magnitude noise, consistently and significantly outperforming traditional DAEs for tabular anomaly detection.
• The optimal number of diffusion steps and choice of noise scheduler are task-dependent: unsupervised settings benefit from more steps and cosine scheduling, while semi-supervised settings prefer moderate steps and linear scheduling.
• Later diffusion steps contribute most to anomaly detection in unsupervised learning, while earlier to moderate steps are more influential in semi-supervised settings due to finer manifold capture.

Summary by alphaXiv.